OpenWorld is an open-source platform for building, fine-tuning, and evaluating robotic policies using video world models. We currently support:
- Fast world model inference with shortcut models
- Policy evaluation on custom benchmarks with automatic scoring
🔧 Some of the functions are still under construction!
Requirements:
- Python 3.11+
- uv for environment management
# Dependencies for base environment only:
uv sync
# Include extra dependencies for using different policies/reward models. Example:
uv sync --extra policy-dp --extra reward-robometer
uv sync --extra policy-openpi --extra reward-robometerFinally, install required assets for the base world model:
sudo apt-get install git-lfs -y
bash external/download_models.sh- 🤖 Policy Training
- 📋 Policy Evaluation
- ⚙️ Policy Fine-tuning (TODO)
- 🌎 World Model Training (TODO)
- Add more reward function support
- Example benchmarks
- Data & checkpoints
This repo is based on Ctrl-World, dppo, dsrl, openpi, and robometer.
Core contributors:
